IP Law Daily, AGENCY NEWS: Speakers at FTC Tech Summit address competitive, consumer risks from rise of AI, (Jan 26, 2024)

By Thomas Long, J.D.

Speakers from industry, academia, government, and media discuss challenges faced by new businesses as well as consumers of AI, as FTC Chair announces inquiry into largest companies in the field.

The explosive surge in generative artificial intelligence (AI) technology over the past 18 months—along with continuing increases in uses of other types of AI and algorithmic decision-making tech—present significant problems with respect to both the promotion of business competition and the protection of consumers from numerous harms, panelists said at the FTC Tech Summit held on January 25. Speakers discussed several aspects of the AI realm—including hardware infrastructure, training of datasets for AI models, and consumer applications—although the key highlight of the event was FTC Chair Lina Khan’s announcement of a sweeping inquiry into the practices of the largest players in this emerging high-tech field.

Remarks of FTC Chair Khan. FTC Chair Lina Khan opened the conference, raising the history of companies that have consolidated market power in their respective industries, including U.S. Steel, Alcoa, and Boeing. In her view, concentrating production also concentrates risk, with recent problems involving Boeing’s 737 MAX airplanes highlighting that fact. Noting the newness of AI technology, Khan said, “We hear how these automated technologies could open up the door to breakthroughs across fields ranging from science to education, making life better for millions of people. But we've also already seen how these tools can turbocharge fraud automate discrimination, and entrench surveillance, putting people in harm's way. More fundamentally, we face basic questions of power and governance. Will this be a moment of opening up markets to fair and free competition, unleashing the full of emerging technologies, or will a handful of dominant firms concentrate control over these key tools, locking us into a future of their choosing?”

Khan stressed that the trajectory taken by AI will be influenced by policy choices. In that connection, Khan delivered the biggest news of the event—the FTC’s launch of an inquiry into investments and partnerships involving generative AI companies and major cloud service providers, aimed specifically at Alphabet, Inc., Amazon.com, Inc., Anthropic PBC, Microsoft Corp., and OpenAI, Inc.

“There’s no AI exception from the laws on the books,” said Khan, “and we’re looking closely at the ways that companies may be using their power to thwart fair competition or trick the public.”

Remarks of Consumer Protection Bureau Director Levine. In remarks delivered at the end of the half-day event, Samuel Levine, Director of the FTC’s Bureau of Consumer Protection, criticized the self-regulatory approach taken during the development of the Internet in the early 2000s, particularly with respect to privacy. According to Levine, by the time the Commission had reversed course, recommending in 2012 that Congress pass privacy legislation, harms to consumers had become entrenched. “This history is very much top of mind as we confront the emerging threats and opportunities created by AI,” Levine said in the written version of his remarks. “Congress has entrusted us with ensuring that markets are fair and competitive, and we will not be sitting on the sidelines.”

Opening the panel discussion, science fiction author Cory Doctorow brought up the issue of copyright problems associated with datasets used to train AI models, such as LLMs. In Doctorow’s views, a licensing scheme—which some have proposed to compensate copyright holders when their works are including in training datasets—would not protect creators because of the market power of a handful of large content owners. In Doctorow’s view, to address harms posed by generative AI would require recourse to labor and privacy law instead of copyright law. He also raised the concern that companies will use AI not just to make workers’ output better but to make it cheaper, with the mistaken expectation that firing half the workforce while adopting AI tools will allow them to get twice the previous output.

First panel: chips and cloud infrastructure. The first group of panelists discussed the competitive landscape in the crucial industries of computer chips and related hardware (including the graphics processing units, or GPUs, relied upon by many AI tools, such as large language models, or LLMs) and cloud computing. Panelists highlighted difficulty faced by new entrants in industries dominated by a small number of large companies, which themselves manufacture their own chips and not just design them while relying on other companies to fabricate them.

Panelist Ganesh Sitaraman of Vanderbilt University pointed to the high degree of concentration throughout the AI “tech stack,” with just three companies dominating the cloud, and fewer than that in chip manufacturing. In Sitaraman’s view, concentration will lead to increasing problems as AI tool creators have few choices to make in obtaining the materials they need for their businesses.

According to Corey Quinn, Chief Cloud Economist at The Duckbill Group, “all roads” in the GPU industry lead through one company—Ndvidia—causing the supply of these expensive components constrained. This creates a risky situation where there is a single point of potential failure. With respect to cloud computing, Reinforcing this view, Quinn said businesses “can’t avoid” the three “hyper-scalers” dominating the industry: Amazon Web Services, Microsoft Azure, and Google Cloud Platform.

Tania Van den Brande is a Director of Economics at Ofcom echoed Quinn’s observation that the competitive landscape is affected by the fact that it is difficult for customers to use more than one cloud computing provider, and expensive to move data from one provider to another. “Competition is not working well in the cloud,” she said. Because companies have economic incentives to use one cloud provider and to stick with that provider, small cloud providers face high barriers to entry. Daven Rauchwerk, a technologist and entrepreneur, also opined that it was difficult for new entrants in the computer chip industry to break in because they “must compete with the largest companies in the world.”

Ganesh added that the possibility of innovation was being foreclosed by vertical integration. “New startups can be out of luck because large vertically integrated companies across the stack can develop their own applications for all the layers,” he said. Ganesh said that a solution could come from a combination of enforcement and regulatory actions, arguing that antitrust enforcement efforts were needed to promote innovation. Chiming in, Quinn said that he didn’t “see a way to fix this,” and Rauchwerk said that he hoped the FTC had some ideas. Ganesh suggested that enforcement options included structural separation of dominant players; promulgation of nondiscrimination rules; regulations to require transparency of terms and conditions; and rules regarding interoperability to facilitate switching cloud providers.

Second panel: data and training. The second panel addressed data and its use in training AI models. Before panelists began their discussion, Commissioner Rebecca Slaughter delivered remarks, expressing the view that we are at a moment of both promise and peril, and suggesting that efforts to promote competition must be taken before it’s too late. Expressing an opinion similar to what speakers said in the first session, Slaughter said that competition in the AI marketplace is hampered by limited access to key inputs controlled by large companies. “Even if a little bit of the hype over the power of AI models is to be believed,” she said, “there is an obligation to support open and fair markets and prevent monopolies in their incipiency.”

Journalist Stephanie Palazzolo opined that startups are deterred from doing anything that would compete with OpenAI. Investors are wary of backing startups that would do anything in areas where OpenAI is doing anything or might do anything, she said. However, a business seeking to develop and market new AI tools doesn’t necessary have to employ the full GPT4 model (referring to OpenAI’s flagship product), but can use more specifically tailored (smaller) models. Palazzolo also urged regulators to press Congress to consider increasing funding of academic research labs as a way to make the playing field less uneven.

Taking a rosier view of the competitive landscape, Jonathan Frankle, Chief Scientist (Neural Networks) at Databricks, highlighted the diversity of products in the AI ecosystem. “Everyone assumes that OpenAI/ChatGPT is the only business model,” he said, but there are many kinds, with numerous businesses creating specialized AI that have existed for decades. Frankle noted that while competition in the cloud industry is intense, new clouds have popped up in the AI world, and there have been new entrants in chip market. However, he opined that competition among businesses is not sufficient to have a good outcome in terms of innovation. He noted that early work in AI was shared in an academic context, but sharing has been shut down because it is “competitive intelligence” that nobody wants to give away. He also pointed to legal and regulatory uncertainty as a challenge, noting that “nobody knows what they’re allowed to use and for what.”

Amba Kak, technology policy expert and Executive Director at the AI Now Institute, expressed the view that data operates as barrier to entry along axes of both quality and scale, explaining that “quality” means data with large amounts of human curation and feedback in the creation of initial datasets. Difficulties are increased by trends toward more restrictions on publicly available data, along with the largest players trending toward less transparency and more opacity regarding their training datasets and techniques. With respect to the tension between AI and copyright, she expressed a similar view to Doctorow, saying that large companies can leverage the existing power dynamic; they can maximize access and push for exclusivity. In addition, with respect to anxieties surrounding cybersecurity, big companies have the edge, the said.

Third panel: Consumer applications. Introducing the third panel, addressing the way AI is being deployed in consumer applications, Commissioner Alvaro Bedoya noted a trend toward LLMs having less transparency around the data being used to create the models. Previously there was a great degree of specificity about what data was used; but now there is little to no transparency. Bedoya emphasized the negative impact this has on the problem of bias in AI outputs, stating that knowing what data is used in training is important for knowing how a model might function. Aside from generative AI systems, there have been problems with discrimination in longstanding uses of algorithmic decision making systems, he said, noting a complaint and settlement over a retail pharmacy chain’s use of a flawed facial recognition system intended to identify shoplifters. “This technology has basic implications for our lives and our ability to lead them,” he said. Companies cannot simply purchase “black box” technology and then claim to lack knowledge about how it works when the tech injures people. Success in regulating AI tech would look like people controlling technology and not the other way around, with people feeling like they’re in control of tech and knowing when it is being used to make decisions about them. In terms of competition, success would mean consumers using a tool not because the biggest company offered it but because it had proven to be the best technology in the marketplace. Startups should have a shot at success, he said.

Panelist Conrad Kramer, co-founder and CTO of Software Applications Inc., said that to be competitive startups need models to build products, which they can train themselves, purchase from other companies, or customize open-source models that are available. The first requires money, expertise, and time, with the second option also being expensive. With open-source models, adopters need to consider possible quality issues, although open-source models offer more transparency than proprietary models with respect to what training data consisted of.

Elaborating on transparency problems, journalist Karen Hao cited the risk this poses to consumers of generative AI systems. She also noted problems caused by ambiguous or deceptive marketing and obfuscation, stating that the way AI developers talk about the tech, they “paper over the hallucination problem,” referring to the tendency for generative AI tools like ChatGPT to sometimes produce output with false or fabricated information. Misleading promotional efforts create a “disconnect between what consumers expect …; and the strengths and weaknesses of what they can get from it,” she said.

Ben Winters, senior counsel for consumer privacy advocacy organization EPIC referred to EPIC’s March 2023 report on harms caused by generative AI, which he said included negative effects on the environment, impacts on elections, individual harms (data security, victims of AI-generated nonconsensual pornography, and others). In his view, misinformation being “supercharged” by the availability of these tools, with the lack of transparency and accountability compounding the problems. With respect to privacy violations occurring in the course of data collection—which increasingly involves intrusive surveillance of individuals, in the view of many—Winters opined that “these systems would not exist if there were good data privacy laws.” Winters also said that it would improve the consumer experience to put the burden for transparency and communication on the companies providing AI tools and not consumers. He also suggested the use of audits and impact assessments as regulatory tools.

Atur Desai, Deputy Chief Technologist for Law and Strategy for the Consumer Financial Protection Board, noted that complex models have been used in financial markets for a long time, despite the current “hype” about novelty. Desai pointed to a robust set of preexisting financial laws, repeating Khan’s warning that “there is no AI or fancy tech exception.” A lack of understanding of tech doesn’t diminish legal obligations regarding lending decisions, he said. If lenders can’t comply with financial disclosure laws while using an AI model, then they shouldn’t use it, in his view.

News: AINews IndustryNewsTrends TechnologyInternet